bugfix (#46921)

python/paddle/distributed/passes/auto_parallel_data_parallel_optimization.py

@@ -82,6 +82,8 @@ class DataParallelOptimizationPass(PassBase):
 
         with paddle.static.program_guard(main_program, startup_program):
             self._analyze_program()
+
+            if self.is_data_parallel_applied():
                 self._prune_grad_scaling()
                 self._calc_comm_overlap()
                 grad_group = self._fuse_allreduce()
@@ -167,6 +169,9 @@ class DataParallelOptimizationPass(PassBase):
         ) == 0, "Unexception: gradients [{}] is scaled BUT NOT synchronized.".format(
             not_synchronized_grads)
 
+    def is_data_parallel_applied(self):
+        return len(self._group_to_grad_name_map) > 0
+
     def _could_be_prune(self):
 
         return self.dist_context.gradient_scale and (